A studio published their own numbers. A 12-minute YouTube video used to eat three to four hours of an editor's attention. It now takes about thirty minutes.
Tooling cost per finished video: forty cents.
That is the claim, and it is from a team running this in production, not a demo reel. They shipped it with the pipeline, the stack, and the parts that still break.
i went in expecting to write about how AI video editors are overhyped. i came out thinking the hype is aimed at the wrong thing.
Claude is not replacing editors. It is sawing the job in half, and one half is much more exposed than the other.
What it actually does
The honest version first. Claude Code cannot watch your video.
It reads the transcript. That is the entire trick, and it is also the entire limitation.
Whisper-class transcription gives you a timecoded transcript of the raw footage at about €0.006 per minute of audio. Claude reads that, proposes cuts with timecoded reasons, and hands back a decision list.
Everything downstream is mechanical. ffmpeg executes the cuts. Remotion renders the motion graphics.
Auphonic masters the audio. ffmpeg encodes the MP4.
The pipeline:
- Ingest the raw file, extract audio, transcribe.
- Propose a cut from the transcript, with a reason per removal.
- Scrub filler words and silences over 1.5 seconds.
- Generate and burn styled captions.
- Place motion graphics from your template library.
- Master loudness to EBU R128.
- Encode.
Six of those seven steps need no human at all. The second one is where you actually spend your time.
An operator accepts 70 to 85 percent of the proposed cuts on the first pass. That number is the whole business case. It is high enough to save hours and low enough that you still have to read every line.
The part nobody says out loud
Filler removal alone accounts for roughly 30 percent of the time savings.
Thirty percent. Not taste, not rhythm, not structure. Just deleting "um" and dead air.
That is the least glamorous possible use of a frontier model, and it is the most valuable one. Because scrubbing ums is the task every editor hates and no client pays for.
the first time i tried this on my own footage, i learned something uncomfortable. My transcript had 41 instances of "so" used as a verbal comma.
i had never noticed. Nobody had ever told me.
Claude noticed in about four seconds.
That is not an editor. That is a mirror with good posture.
Templates are the moat, not the code
here's a question people always ask about this pipeline: what do i actually need to build?
The answer is not the orchestration. That is roughly 200 lines of code, and Claude can write most of it for you.
The answer is eight to twelve motion-graphic templates.
Look at your last ten videos. Find the visual elements that repeat. The title card, the lower-third layout, the callout treatment, the stat highlight, the end card.
Build those as Remotion components, because React compiles to video and lives in git like anything else.
The generic template library produces generic videos. A custom set of eight produces a channel that looks consistent across a hundred episodes.
The team that published this called the templates the moat and the code commodity. i think they are right, and i think most people will get this backwards.
What breaks, every time
Claude cannot do the opening eight seconds.
It can suggest four candidate openings. It cannot pick the right one, because the decision depends on whether the first line makes a stranger stop scrolling, and that is not a property of text.
It also cannot spot the moment you should keep.
The awkward two-second pause before the joke lands. The off-script aside where the speaker corrects themselves. The take that looks worse on paper and is the only one that works.
AI's bias is to clean these up. The right call is often to leave the mess in.
And it cannot feel the pacing of a back half. Cuts are transcript logic. Rhythm requires watching the result and noticing that minute nine is sagging.
| Claude does this | You do this |
|---|---|
| Filler and dead air | The opening eight seconds |
| Caption burn-in | Which awkward pause stays |
| Motion graphic placement | Pacing of the back half |
| Loudness and encode | Knowing when to break the format |
One column is mechanical. The other is why anybody hires a person.
A detour about watching your own raw footage
Nobody warns you about this part.
Watching an hour of yourself talking, unedited, is a specific kind of unpleasant. You notice the blink rate. You notice that you say "right?" at the end of every third sentence.
Then you notice the moment your eyes drift to the second monitor and your whole face goes blank for two seconds.
Editors see this all day and never mention it. That is a professional courtesy i did not know existed until i did it to myself.
The pipeline makes it easier, and it does not make it less strange. Handing raw footage to a machine that will read every word you said and rank which ones were filler is a new kind of exposure.
i have started recording one fewer take. Not because of quality. Because i know someone is going to read the transcript.
Who should not build this
If you publish one video a month, do not build a pipeline. The setup cost is a long weekend, plus another weekend to redo the parts you got wrong, and you will spend more time maintaining it than you save.
If your videos are under five minutes, skip it. The mechanical overhead does not amortize.
If you are editing for a client who pays by the hour, this is not a productivity win. It is a pricing problem, and you should think about it as one.
And if your work is mostly taste, the pipeline does nothing for you. Wedding films. Documentary.
Anything where the cut is the argument.
Here is the uncomfortable part for freelancers. The editor charging €40 to €80 an hour for cuts and captions is competing, whether they know it or not, against a pipeline that does 80 percent of that deliverable for forty cents.
The senior editor at €120 to €200 an hour is fine. Their buyer was never paying for cuts.
The middle is where it hurts. Generalists doing both halves at €60 to €100 now have to pick a side.
The bit that stuck
The team behind this did not describe a tool. They described a role that should not exist anymore: editor as cutter.
And a role that suddenly makes sense: editor as director. Someone who never opens a timeline. Who watches the raw footage, writes a cut spec, hands it to a junior running the pipeline, and approves the result.
They quoted their head of operations on a freelancer who made that pivot in two months. Same person. Same taste.
Moved from per-video cutting to a monthly retainer, and stopped selling hours entirely.
The pipeline did not take that job. It made the job visible.
Forty cents of tooling sounds like the end of something. Mostly it is the end of the part nobody wanted to do.
Top comments (1)
The point about AI cleaning up the parts that should actually stay is interesting. Removing filler words and silences sounds great until you realise that sometimes the awkward pause or an off-script moment is what makes a video feel real. I think the bigger shift here is that editors might spend less time cutting footage and more time deciding what deserves to stay.